1. The SNOW Theorem and Latency-Optimal Read-Only Transactions
Haonan Lu❄,
Christopher Hodsdon❄, Khiem Ngo❄,
Shuai Mu†, Wyatt Lloyd❄
❄University of Southern California, †New York University

2. Huge Web Services Shard Data
Massive amount of data
 must be distributed across servers
Reads dominate the workloads
– need to be as fast as possible!

3. Simple Reads Are Insufficient
Datacenter
Web
Storage Tier
Load
Page
Clients
Servers
Read
Public
ACL=private
ACL=public
Read
Photo B
Set
“Private”
Photo B
Done
Not Acceptable!
Private
Photo B
Done
Photo B
Done
Photo=A
Photo=B

4. Read-Only Transactions
Transactions that do not modify data
Consistently read data across servers

5. The Power of Read-Only Txn
Consistency restricts what can be read
Eliminates unacceptable combinations
Compatibility enables write transactions
Write transactions atomically update data
Higher power  more useful
Stronger consistency  higher power
Compatibility  higher power

6. highest power + lowest latency is impossible
Intuitive Tension
Fundamental Tradeoff
High Power
Low Latency
Reduces anomalies (the ACL – Photo example)
Better user experience
Easier to reason about
Higher revenue
Our study proves:
highest power + lowest latency is impossible

7. The SNOW Properties [S]trict serializability [N]on-blocking operations
[O]ne response per read
[W]rite transactions that conflict
Highest
Power
Lowest
Latency

8. [S]trict Serializability
Strongest model: real-time + total order CR
SACL
SPhoto CW
ACL := Private
Upload Photo B
W starts
Private
“Photo B is private!”
Photo B
W finishes
R starts
R finishes

9. [S]trict Serializability
Strongest model: real-time + total order CR
SACL
SPhoto CW
W starts
R starts
Photo B
ACL := Private
Upload Photo B
“Public + Photo A”
“Photo B is private!”
Private
“Public + Photo B”
“Photo A is private!”
W finishes
R finishes

10. [N]on-blocking Operations
Do not wait on external events
Locks, timeouts, messages, etc.
Lower latency
Save the time spent blocking

11. [O]ne Response One round-trip One value per response
No message redirection
Centralized components: coordinator, etc.
No retries
Save the time for extra round-trips
One value per response
Less time for transmitting, marshaling, etc.

12. [W]rite Transactions That Conflict
Compatible with write transactions
Richer system model
Easier to program

13. The SNOW Theorem: Impossible for read-only transaction
algorithms to have all SNOW properties

14. Why SNOW Is Impossible CW SA SB CR W starts A := new B := new R
Assume
SNOW  R W
invisible
Violates
property S T W
visible
RA = new
RB = old
W finishes

15. A Deeper Look at SNOW Complete proof in the paper SNOW is tight
Any combination of 3 properties is possible
Optimality
SNOW-optimal: have any 3 properties
Latency-optimal: have property N and O
Spectrums of property S and O
Show what is possible to achieve

16. Study Existing Systems with SNOW SNOW-optimal and latency-optimal
Spanner-Snap
[OSDI’12] ✖ ✔
Yesquel
[SOSP’15]
MySQL Cluster
Spanner-Snap
[OSDI ’12] ✖ ✔ ✔ ✔
Yesquel
[SOSP ’15] ✖ ✔ ✔ ✔
MySQL Cluster ✖ ✔ ✔ ✔

17. Study Existing Systems with SNOW SNOW-optimal
Eiger [NSDI ’13] ✔
≤ 3
DrTM [SOSP ’15]
≥ 1
RIFL [SOSP ’15]
≥ 2
Sinfonia [SOSP ’07]
Spanner-RO [OSDI ’12] ✖ ✔ ✖ ✔ ✔

18. Study Existing Systems with SNOW Candidates for Improvement
COPS ✖ ✔
≤ 2
Rococo
> 1
Many more

19. Improve Existing Systems with the SNOW Theorem
COPS [SOSP ’11]
Geo-replicated
Causally consistent
Read-only txn
Rococo [OSDI ’14]
Supports general transactions
Strictly serializable
: S N O W
: S N O W

20. New Algorithm Designs COPS-SNOW Rococo-SNOW
Latency-optimal (N + O)
Rococo-SNOW
SNOW-optimal (S + O + W)
Design insight for optimizing reads:
shift the overhead to writes

21. Rococo’s Read-Only Txn (S + W)CR SA SB CW
W starts
A := “new”
B := “new”
R: 1st round
Gather
conflict info
A=“old”
EQUAL ?
Blocks
W commits
B=“new”
R: 2nd round
W finishes
R: Nth round

22. Strictly Serializable
Rococo-SNOW (S+O+W) CR SA SB CW
W starts R TS
A := “new”
B := “new” TS
A=“old”
Strictly Serializable
Forward TS
Blocks
W commits TS
B=“old”
A=old
B=old
W finishes

23. Evaluation of Rococo-SNOW
To understand
Latency of read-only transactions
Throughput of other types of transactions
Experiment configuration
Identical to Rococo’s
TPC-C workloads

24. Significantly Lower Latency for Read-Only Txn
OCC
Lock Wait
Retries
Always 1 round
Rococo
2PL
High Contention
Rococo
-SNOW

25. Higher Throughput under High Contention
Rococo
-SNOW
Rococo
-14% throughput
(Low Contention)
2X throughput
(High Contention)
High Contention
2PL
OCC

26. Conclusion The SNOW Theorem for read-only txns
Impossible to have all of the SNOW properties
SNOW helps understand existing systems
Many are not yet optimal
Rococo-SNOW
SNOW Theorem guided SNOW-optimal design
Significantly higher throughput and lower latency under high contention